Semi-supervised Learning for Myopic Maculopathy Analysis

Myopia is a common ocular disease that affects large populations in the world. This disease can lead to visual impairment due to the development of different types of myopic maculopathy. Therefore, prompt screening and intervention are necessary to prevent the further progression of myopic maculopathy to avoid vision loss. However, the manual inspection process is a time-consuming task that relies heavily on the experience of ophthalmologists. Towards advancing the state-of-the-art in automatic myopic maculopathy analysis using retinal fundus images, the “Myopic Maculopathy Analysis Challenge (MMAC) 2023” was launched. In this work, we present how semi-supervised learning methods can be applied to tackle two of the tasks proposed in the MMAC challenge: the segmentation of myopic maculopathy plus lesions, and the prediction of spherical equivalent. In particular, we have applied a pseudo-labeling approach for the segmentation task obtaining an average Dice similarity score of 0.6706, and a data-distillation procedure for the regression problem obtaining an R-square score of 0.7906. In both cases, we used the data from the other tasks of the challenge to increase the size of the original dataset with an automatically annotated dataset. This approach obtained the fourth position in both the segmentation and spherical equivalent prediction tasks.